Look Ahead – Machine Learning in Radiology
Given the many scientific papers, newspaper and magazine articles, and even television ads about machine learning (ML), big data, data mining and artificial intelligence (AI), it is not surprising that RSNA News is featuring a Look Ahead article on this topic. After all, how can radiologists ignore IBM’s Watson reading an X-ray during half-time of an NFL game? While I’m not a computer scientist, I suspect that the radiology community is more interested in the clinical practice implications of this technology than the technology itself. Having practiced radiology and managed clinical practices for over 40 years, I will build on that experience to discuss ML and its role in tomorrow’s radiology practice.
What is Machine Learning?
Not unexpectedly, radiologists and computer scientists think about ML quite differently. The radiologist might ask,“Can a machine learn to do what we, as radiologists, do?” An important corollary question is, “Can we teach a machine what we know and to do what we do?” Yet the central question of contemporary ML is, “Can a machine learn more than what we now know and use this new knowledge to make decisions?”
To a computer scientist, ML means much more than programming, or “teaching” a computer to perform a specific, known task. In 1959, Arthur Samuel, a pioneer in ML and AI, defined machine learning as the “field of study that gives computers the ability to learn without being explicitly programmed.” Machine learning implies algorithms that can learn from and make predictions based on new data.
Implications for Radiology
In general, ML algorithms create computational models based on example inputs in order to derive data-driven predictions or decisions. In radiology, the main decision-making task performed is one of classification— given the images and available clinical information, what is the most likely diagnosis? Innumerable machine learning algorithms perform such classification tasks, though they differ in specific implementation, mathematical basis and logical organization. While knowledge of the task (diagnostic) performance of these algorithms is critical to the practicing radiologist, knowledge of programming details is not necessary. Whether an algorithm utilizes neural networks, support vector machines, clustering analysis, Bayesian networks, sparse coding, genetic algorithms, random fields, or ‘deep learning,’ its value to the radiologist is dependent on how accurately a given application makes diagnoses.
Fortunately ML algorithm performance can be documented through statistical metrics with which radiologists are familiar, such as receiver operating characteristic (ROC) analysis. When the area under the curve (AUC) of an ML algorithm applied to a radiologic task approaches that of a radiologist, the machine will have learned something we now know and will be able to make decisions based on that knowledge. At that point, consideration should be given to incorporating that algorithm into clinical practice. When, and if, the AUC of an ML algorithm exceeds that of a radiologist, the machine will have learned more than what we now know and will be able to make decisions we cannot now make. At that point, we may be compelled to incorporate that algorithm into our practice. Incorporating these analytical tools into our practice, however, does not mean replacing radiologists. Rather, they will complement our quite remarkable human skills
“Supervised” vs. “Unsupervised” Learning
As with humans, ML algorithms can “learn” through different processes. A useful, if overly simple, dichotomy in ML processes is supervised versus unsupervised learning. In the former, example inputs (chest X-rays) and desired outputs (the chest X-rays’ respective diagnoses: normal/lung cancer) are given to the computer by a supervisor (or teacher) and the computer learns general rules that discriminate between the possible diagnoses. Simplistically, the machine learns from what is known. We do the same thing with human trainees by showing them diagnosis-proven cases. Both ML algorithms and radiologists learn better (i.e., increase their diagnostic accuracy) by increasing the number of training cases, thousands of which are needed to train a human or a computer. Hence the need for long residency/fellowship training or big data in radiology.
In unsupervised learning, the inputs (chest X-rays) might be the same, but no labeled outputs are provided. The computer has to figure out what the possible diagnoses are and how to discriminate amongst them. In this model, the machine is trying to learn from what is unknown in the data. This leads toward data mining and knowledge discovery. Supervised learning is often more expedient and takes advantage of our extensive prior knowledge of radiologic diagnoses, but is obviously limited in finding new patterns of disease. Academic radiologists participate in unsupervised learning when we are searching for a new disease pattern, which is a harder task and often takes more input cases than learning and making a known diagnosis. There is a semi-supervised middle ground and the earliest ML algorithms used May in radiology will almost certainly be based on supervised or semi-supervised learning technology, with less supervised data mining technology operating on even bigger data sets possibly yielding new diagnostic capabilities.
Identifying Patterns in Big Data
Intimately related to ML, particularly in relationship to image data, is the concept of patterns—a set of observations, ideally measurements, of any object. In the case of radiologic images, the most basic pattern is that of the relative intensity (reflected by brightness or darkness) of every signal (e.g., radiodensity, T1, echogenicity) measured for every spatial pixel or voxel, at every time point of a set of images. Those measurements comprise the raw data of an image and result in a complex pattern. Even with relatively low-resolution, single-modality examinations, image data fits into the “large data” category (e.g., non-contrast enhanced CT scan of the head [1.0 to 50 MB]); with higher resolution, multi-dimensional studies being big data (e.g., coronary CTA [200 MB to 2 GB]); and the patterns within population studies such as the Alzheimer Disease Neuroimaging Initiative (multi-modality brain MRI scans with DTI and fMRI [TB]) basically defying human comprehension.
The human visual system is remarkable and radiologists learn and remember very complex patterns and use them every day to make clinical decisions and diagnoses. Modern imaging technology, however, is creating image data sets that exceed human pattern recognition capabilities. Computers and ML technology feast on such data and are rapidly becoming capable of learning incredibly complex, multi-dimensional patterns derived from large normal and diseased populations. That data may be used to diagnose known diseases, such as Alzheimer’s disease, but potentially could also define new patterns for diseases such as schizophrenia.
Machine Learning as an Emerging Tool
Many ML algorithms are now being developed for specific radiologic diagnosis with ROC AUCs exceeding that of house staff and approaching that of board-certified radiologists. These algorithms are relatively low level and/or narrowly task specific. This piecemeal development of ML applications takes advantage of subspecialty expertise in building disease and task models. ML programs will expand into clinical practice in a piecewise fashion, at first meeting small niche tasks, but eventually coalescing into more powerful, broadly applicable diagnostic tools.
The paucity of ML tools in radiology will rapidly change with the increased ability of ML algorithms to participate in radiology’s most critical decision processes—the detection and diagnosis of disease. For instance, ML algorithms incorporated into computer assisted detection/diagnosis (CADD) products are now detecting pulmonary nodules, diagnosing colonic polyps and screening for breast cancer, with much more to come. In the not so far future, ML will play a central role in radiology, becoming part of routine workflow and providing daily real-time clinical diagnostic support. I predict that within 10 years no medical imaging study will be reviewed by a radiologist until it has been pre-analyzed by a machine.
But I do not foresee Watson taking over our jobs. The human visual system and brain have extraordinary abilities to work with incredibly noisy data, assimilate disparate information, intuit unexpected insights, conjure alternative scenarios and solve incredibly complex problems that are common in medicine, but well beyond any known computer solutions. Much of life and radiology is not that complicated, however, and in that more simple space is where technological advances first occur. Many simple tasks radiologists perform can and will be performed by computers, and ML will expand these capabilities into our central task of image interpretation.
Machine Learning Improves Quality, Efficiency
The main factors driving change in radiology are the demand for improved quality and greater efficiency. ML technology can reduce human errors and more rigorously analyze our increasingly complex imaging examinations. The demand for imaging services continues to increase while reimbursement rates decrease. To meet demand and maintain compensation, radiologists already read at least 20 percent more cases per day than they did 10 years ago and view more than twice as many images. In terms of human visual psychometrics, we are operating at near capacity. Further gains in clinical efficiency will require new technologies such as ML algorithms that preview images and create draft reports for the radiologist. Academic radiology departments are currently evaluating such prototype systems. Cleverly integrated into imaging workflow and with appropriate human oversight, ML technology will allow the radiologist to work more efficiently and produce better quality reports.
http://www.rsna.org/News.aspx?id=19018 | May 1, 2016 |